Association Rules and Refined Information Granulation-Based Time-Series Long-Term Forecasting

Accurate long-term prediction is a crucial and challenging task in time series analysis. The granular modeling, which can capture time series patterns and trends by transforming them into information granules, offers a promising strategy for long-term prediction. This study designs a long-term forecasting method by refining the granular representation of time series and identifying significant hidden relationships within them. First, following the principle of justifiable granularity, a series of trend-based information granules are constructed, and with the incorporation of patch learning, the granular results are refined to capture the essential characteristics of the time series as fully as possible. After segmenting at a fixed time granularity, multi-trend fuzzy association rules ( multi-trend FARs) are developed to analyze the situation when multiple patterns emerge in a time granularity, where short and long perspectives are both incorporated to extract the corresponding relationships. Further, with the aid of improved fuzzy C-means, a rule selection scheme is designed to determine the available multi-trend FARs and implement the prediction results. Here, an optimization of the weights for short and long FARs is developed to ensure the prediction accuracy. Finally, experiments conducted on several datasets show that the proposed method achieves higher accuracy and justifiability compared with existing models across different prediction horizons, indicating the practicality and efficacy of the proposed forecasting method in long-term prediction.